Crop pest classification with a genetic algorithm-based weighted ensemble of deep convolutional neural networks

Abstract

Insects are among the important causes of significant losses in crops such as rice, wheat, corn, soybeans, sugarcane, chickpeas, potatoes. Identification of insect species in the early period is crucial so that the necessary precautions can be taken to keep losses at a low level. However, accurate identification of various types of crop insects is a challenging task for the farmers due to the similarities among insect species and also their lack of knowledge. To address this problem, computerized methods, especially based on Convolutional Neural Networks (CNNs), can be employed. CNNs have been used successfully in many image classification problems due to their ability to learn data-dependent features automatically from the data.Throughout the study, seven different pre-trained CNN models (VGG-16, VGG-19, ResNet-50, Inception-V3, Xception, MobileNet, SqueezeNet) were modified and re-trained using appropriate transfer learning and fine-tuning strategies on publicly available D0 dataset with 40 classes. Later, the top three best performing CNN models, Inception-V3, Xception, and MobileNet, were ensembled via sum of maximum probabilities strategy to increase the classification performance, the model was named SMPEnsemble. After that, these models were ensembled using weighted voting. The weights were determined by the genetic algorithm that takes the success rate and predictive stability of three CNN models into account, the model was named GAEnsemble.GAEnsemble achieved the highest classification accuracy of 98.81% for D0 dataset. For the sake of robustness ensembled model, without changing the initial best performing CNN models on D0, the process was repeated by using two more datasets such that SMALL dataset with 10 classes and IP102 dataset with 102 classes. The accuracy values for GAEnsemble are 95.15% for SMALL dataset and 67.13% for IP102. In terms of performance metrics, GAEnsemble is competitive compared to the literature for each of these three datasets.